Health recommender system design in the context of CAREGIVERSPRO-MMD project

Heal h Recommende Sys em design in he con ex o CAREGIVERSPRO-MMD P ojec Luis Oli a-Felipe Uni e si a Poli ècnica de Ca alunya - Ba celonaTech Ba celona, Ca alunya, Spain [email p o ec ed] C is ian Ba ué Uni e si a Poli ècnica de Ca alunya - Ba celonaTech Ba celona, Ca alunya, Spain [email p o ec ed] A ia Co és Uni e si a Poli ècnica de Ca alunya - Ba celonaTech Ba celona, Ca alunya, Spain [email p o ec ed] Emma Wol e son Uni e si y o Hull Hull, Uni ed Kingdom e.w[email p o ec ed] Ma co An oma ini COOSS Ma che ONLUS Ancona, I aly [email p o ec ed] Isabelle Land in Cen e Hospi alie Uni e si ai e de Rouen Rouen, F ance Isabelle.Land in@chu- ouen. Kons an inos Vo is In o ma ion Technologies Ins i u e Cen e o Resea ch and Technology Hellas Thessaloniki, G eece [email p o ec ed] Ioannis Paliokas In o ma ion Technologies Ins i u e Cen e o Resea ch and Technology Hellas Thessaloniki, G eece [email p o ec ed] Ulises Co és Uni e si a Poli ècnica de Ca alunya - Ba celonaTech Ba celona, Ca alunya, Spain [email p o ec ed] ABSTRACT CAREGIVERSPRO-MMD an EU H2020 unded p ojec aims o build a digi al pla o m ocusing on people li ing wi h demen ia and hei ca egi e s, o e ing a selec ion o ad anced, indi idually ailo ed se ices enabling hem o li e well in he communi y o as long as possible. This pape p o ides an ou line o a heal h ecommende sys em designed in he con ex o he p ojec o p o ide ailo ed in e en ions o ca egi e s and people li ing wi h demen ia. CCS CONCEPTS •Human-cen e ed compu ing →Human compu e in e ac- ion (HCI);Sys ems and ools o in e ac ion design; KEYWORDS Recommende sys ems, Social Ne wo ks, Demen ia, Ca egi e ACM Re e ence Fo ma : Luis Oli a-Felipe, C is ian Ba ué, A ia Co és, Emma Wol e son, Ma co An- oma ini, Isabelle Land in, Kons an inos Vo is, Ioannis Paliokas, and Ulises Co és. 2018. Heal h Recommende Sys em design in he con ex o CAREGIVERSPRO- MMD P ojec . In PETRA ’18: The 11 h PE asi e Technologies Rela ed o As- sis i e En i onmen s Con e ence, June 26–29, 2018, Co u, G eece. ACM, New Yo k, NY, USA, 8 pages. h ps://doi.o g/10.1145/3197768.3201558 ACM acknowledges ha his con ibu ion was au ho ed o co-au ho ed by an employee, con ac o o a ilia e o a na ional go e nmen . As such, he Go e nmen e ains a nonexclusi e, oyal y- ee igh o publish o ep oduce his a icle, o o allow o he s o do so, o Go e nmen pu poses only. PETRA ’18, June 26–29, 2018, Co u, G eece ©2018 Associa ion o Compu ing Machine y. ACM ISBN 978-1-4503-6390-7/18/06...$15.00 h ps://doi.o g/10.1145/3197768.3201558 1 INTRODUCTION Acco ding o he Wo ld Heal h O ganisa ion [ 31 ], 47M people a ound he wo ld ha e some o m o demen ia, o which he e is no e ec i e in e en ion, o hal o e e se he p og essi e cogni- i e impai men . As Eu ope’s popula ion is ageing, long- e m ca e o elde ly ci izens will become an inc easing cos o socie y. To manage his ansi ion, heal hca e policies in he EU and indi idual Membe S a es a e hea ily ocused on ex ending he independen li e o he elde ly, wi h he dual aim o inc easing hei quali y o li e and educing he cos s o ca e. Heal h deli e y p ac ices a e shi ing owa ds home ca e. The easons a e he be e possibili ies o managing ch onic ca e, con- olling heal h deli e y cos s, inc easing quali y o li e and quali y o heal h se ices and he dis inc possibili y o p edic ing and hus a oiding se ious complica ions [ 17 ]. The concep o social ne wo ks using mobile de ices o suppo heal communi ies has eme ged as a new and iable eali y in he ield o IT o heal h and elemedicine unde he unding o he EU p og ams. The so-called mHeal h has he “po en ial o ans o m he ace o heal h se ice deli e y ac oss he globe", and he e is inc easing media and consume in e es in mHeal h, illus a ed o example by a ecen panel a he US Con- sume Elec onics Show on The Digi al Heal h Re olu ion [ 19 ],[ 29 ], [30]. Ageing is one o he g ea es social and economic challenges o he 21 s cen u y o Wo ld socie ies. I will a ec all EU coun ies and mos policy a eas. The Ageing o Eu ope is cha ac e ised by a dec ease in e ili y, a dec ease in mo ali y a e, and a highe li e expec ancy among Eu opean popula ions, oge he wi h con inued bu decele a ing inwa d ne mig a ion o he EU. In all Eu ope, li e expec ancy is inc easing in an almos con inuous and uni o m end a he a e o 2-3 mon hs e e y yea and is he main d i e behind he popula ion ageing. By 2025 mo e han 20% o Eu opeans PETRA ’18, June 26–29, 2018, Co u, G eece L. Oli a e al. will be 65 o o e , wi h a pa icula ly apid inc ease in numbe s o o e -80s. Due o an ageing popula ion, he public p o ision o long- e m ca e poses an inc easing challenge o he sus ainabili y o public inances in he EU. The ageing o he popula ion is expec ed o pu p essu e on go e nmen s o p o ide long- e m ca e se ices as old people o en de elop mul i-mo bidi y condi ions, which equi e long- e m medical ca e and assis ance wi h a numbe o daily asks. Heal h ends among he elde ly a e mixed: se e e disabili y is declining in some coun ies bu inc easing in o he s, while mild dis- abili y and ch onic disease a e gene ally inc easing. This si ua ion is compa able in many socie ies a ound he wo ld. CAREGIVERSPRO-MMD (C-MMD) is an EU unded p ojec unde he H2020 p og amme de o ed o building a mHeal h applica ion ha is speci ically a ge ed o ca egi e s and people wi h cogni i e impai men o mild o mode a e demen ia. The no el y o C-MMD consis s in o in eg a ing a b oade diagnos ic app oach, inco po a - ing he li e-in amily Ca egi e -Pe son li ing wi h demen ia dyad and conside ing his dyad as he uni o ca e. CAREGIVERSPRO-MMD is ocused on people li ing wi h Mild Cogni i e Impai men o Mild o Mode a e Demen ia (PLWD om now on) and hei ca egi e s. Mild Cogni i e Impai men (MCI) is a condi ion ha alls somewhe e be ween no mal age- ela ed memo y loss and Alzheime ’s disease o a simila impai men . No e e yone wi h MCI de elops demen ia. And like demen ia, MCI is no an illness, bu a clus e o symp oms ha desc ibe changes in how you hink o p ocess in o ma ion. Memo y p oblems a e he mos common indica o s o MCI. A pe son wi h MCI may also expe ience di icul ies wi h judgemen , o ien a ion, hinking and language beyond wha one migh expec wi h no mal ageing. Fo unknown easons, MCI appea s o a ec men mo e han women. The p ojec comp ises h ee phases: i s , o design and de elop he i s p o o ype o he mHeal h applica ion conside ing p e i- ous wo k on demen ia and psychia ic co-mo bidi y symp oms, sc eening and in e en ion s a egies o be implemen ed by he pla o m. In he second phase, o conduc a use -cen ic analysis o e- design he exis ing p o o ype o PLWDs. The de elopmen was s ee ed by PLWDs, ca egi e s and doc o s, h ough use -cen ic design: eedback is being collec ed on each new e sion o he applica ion un il he design is adap ed o he use s’ condi ions. In he hi d phase, he op imised applica ion is being pilo ed wi h 600 dyads (PLWDs and hei espec i e ca egi e s) and 600 con- ols. This will show he clinical and social bene i s o PLWDs and ca egi e s, as well as inancial bene i s o he heal hca e sys em. C-MMD is an in elligen suppo pla o m p omo ing Quali y o Li e (QoL), well-being and medica ion compliance o PLWD and Ca egi e s in he communi y a he poin o ca e. I will be a ailable on sma phones and able compu e and web b owse s. I s in e ace is being designed o sui use s wi h low IT amilia i y [ 20 ]. Addi ionally, C-MMD will be complian wi h in e nal secu i y p o ocols and policies as well as indus y egula o y policies, i is designed o only collec and p ocess da a conce ning heal h o speci ic and legi ima e pu poses. Some o he needs ha PLWD and hei in o mal ca egi e s cu en ly pe cei e as insu icien ly me by egula ca e and suppo se ices migh be alle ia ed, o e en be me wi h he help o he mu ual assis ance communi ies, using C-MMD. These needs can be summa ized as (i) he need o gene al and pe sonalized in o ma- ion; (ii) he need o suppo wi h ega d o symp oms o demen ia; (iii) he need o social con ac and company; and (i ) he need o heal h moni o ing and pe cei ed sa e y [14]. The he apeu ic educa ional in e en ions is ano he C-MMD’s impo an se ice. This se ice o e s pe sonalized educa ional con- en s o hei PLWDs and ca egi e s h ough he social ne wo k eed ailo ed o hei speci ic p o ile. The educa ional ocus is wo- old: i s , i o e s gene ic con en s ocused on demen ia disease, demen ia and psychia ic co-mo bidi y symp oms and a ailable esou ces on he communi y and secondly, i p oposes use - ailo ed con en igge ed on he basis o use ’s p o ile. In his pape , we will ocus on his se ice p o ision. 1.1 Plan o he wo k The plan o his pape is he ollowing, in §2 we will explain he concep o in e en ion and he speci ic s a egy selec ed in C-MMD. Sec ion 3 p o ides backg ound in ecommende sys ems. In §4 we explain he ecommende sys em app oach de eloped in C-MMD. In §5 we will discuss ou conclusions and commen he u u e wo k. 2 PSYCHOSOCIAL INTERVENTIONS A heal h in e en ion is an ac pe o med o , wi h o on behal o a pe son o a popula ion whose pu pose is o assess, imp o e, main- ain, p omo e o modi y heal h, unc ioning o heal h condi ions [ 22 ]. A b oad ange o p o ide s can ca y ou in e en ions ac oss he ull scope o heal h and social sys ems including acu e ca e, p ima y ca e, ehabili a ion, and assis ance wi h unc ioning, p e- en ion and public heal h. Fu he de ini ions conside Educa ional In e en ion as an in e en ion ha aims o educa e, in o m and shape unde s anding o demen ia and ca egi ing p ac ice [ 9 ] and Psychosocial in e en ion as a b oad e m used o desc ibe di e en ways o suppo people o o e come challenges and main ain good men al heal h [ 7 , 8 , 10 , 23 ]. In he li e a u e, o en he e ms psy- chosocial and nonpha macological a e used synonymously o e e in e en ions. Psychosocial in e en ions a e o en adminis e ed by a ained pe son, e.g. a GP, a psychologis o psycho he apis , occupa ional he apis s o nu ses, and a e usually pe son-cen e ed. These so o in e en ions a e expec ed o help PLWD and o en hei ca egi e s as well, wi h [10]: •coming o e ms wi h a diagnosis o demen ia •main aining social li e and ela ionships a e diagnosis • educing s ess and imp o ing mood, anxie y o dep ession • educing o adap ing beha iou diso de s • imp o ing cogni i e unc ions, such as hinking and memo y •li ing independen ly • main aining and imp o ing quali y o li e - main aining heal h and happiness, and con ol o e one’s li e •suppo ing he pa ne and amily His o ically, mos psychosocial in e en ions ha e been adminis- e ed in a ace- o- ace o ma be ween a p o ide (a ained pe son) and a ca e ecei e (a PLWD, a Ca egi e o bo h). Mo e ecen ly, hese ha e also included he use o elephones o o he digi al de- ices, ideo con e ences, sel -guided books o In e ne ideos [ 21 ]. Some in e en ions combine one o mo e o hese op ions. In he Heal h Recommende Sys em design in he con ex o CAREGIVERSPRO-MMD P ojec PETRA ’18, June 26–29, 2018, Co u, G eece scope o C-MMD pla o m, he ocus is on non-pha macological educa ional in e en ions o PLWD and hei ca egi e s ha can be p o ided h ough digi al means. Tailo ed in e en ions, which a e pe sonalized o he indi idual and assessmen based, ake he concep one s ep u he by adap ing he in e en ion o speci ically ma ch he indi idual pa ien p o ile. This wo k desc ibes he echnical app oach aken o implemen he pe sonaliza ion o he con en s p esen ed o he C-MMD pla o m use s. 2.1 C-MMD In e en ions In e en ions a e well desc ibed in he li e a u e bu he e is no widely accep ed s anda d classi ica ion sys em ha would allow o a mo e comp ehensi e unde s anding o he di e en ypes o ea men s and ca e p ac ices, he cha ac e is ics o people who bene i om hem, o he kinds o p oblems each one has been shown o educe o esol e [ 2 ]. Howe e , many o ganisa ions p o- ide hei own p oposals unde di e en c i e ia. The NICE-SCIE, o ins ance, ca ego izes in e en ions acco ding o hei use: cogni- i e symp oms, non-cogni i e symp oms and beha iou , como bid emo ional condi ions and ca egi e suppo . On he o he hand, o he classi ica ions can also iden i y in e en ions o he p ob- lem hey a ge (e.g. sleep p oblems). In he C-MMD ecosys em wo co e g oups o s akeholde s can be ound, di ided hei knowl- edge backg ound: (i) Heal h and Social P o essionals (HSP) and (ii) Ca egi e s/PLWD. The impo ance o de eloping and using language and e minology ha is unde s andable and meaning ul o hese s akeholde g oups in all e o s o expand he a ailabil- i y and use o non-pha macological ea men s and ca e p ac ices mus be emphasized. The C-MMD pla o m mus allow he HSP o classi y co ec ly he in e en ions and con en s hey c ea e o he PLWD and Ca egi e s while hese mus be able o iden i y hei p oblems and needs in he ca ego isa ion. The C-MMD Conso ium has explo ed se e al in e en ion axonomies om us ed sou ces like: • In e dem, a pan-Eu opean ne wo k o esea che s collabo- a ing in esea ch on and dissemina ion o Ea ly, Timely and Quali y Psychosocial In e en ions in Demen ia [18] • The Na ional Ins i u e o Heal h and Clinical Excellence - Social Ca e Ins i u e o Excellence Guideline on Suppo ing People Wi h Demen ia and Thei Ca e s in Heal h and Social Ca e [9] • The B i ish Psychology Socie y Guide o Psychosocial In e - en ions in Ea ly S ages o Demen ia [10] • Demen ia Ca e si e de eloped wi h and by amily membe s who look a e someone who has demen ia [28] The C-MMD Conso ium has iden i ied he domains o he De- men ia Ca e p oposal as he mo e signi ican , as hey e lec he mos he poin o iew o dyad daily li e bu i mus be no ed ha he e is a consis en pa allelism in he di e en p oposals. Based on li e a u e su ey, he C-MMD Conso ium depic ed a lis o 26 in e en ions ca ego ized by in e en ion a ea om which 12 we e selec ed (see Table 1) o be as candida es o be in eg a ed in o he C- MMD pla o m upon he easibili y o be echnically implemen ed. 2.2 In e en ion P o ision Acco ding o NICE-SCIE, ea men and ca e should ake in o accoun pa ien s’ needs and p e e ences (o heal hca e goals). People wi h de- men ia should ha e he oppo uni y o make in o med decisions abou hei ca e and ea men , in pa ne ship wi h hei heal hca e p o- essionals. One o he main cha ac e is ics o he C-MMD pla o m is he possibili y o empowe bo h he PLWD and he Ca egi e s imp o ing he abili ies o he sel -managemen o hei heal h and aking in o accoun hei a i udes and inclina ion. Fo such eason, he s a egy o in e en ion p o ision is also es ablished acco ding o he use ’s p e e ences. One o he C-MMD p ima y se ices is o p o ide use - ailo ed in e en ions, anked acco ding o use s’ needs and p e e ences and sc eening esul s. In o de o p o ide ailo ed in e en ions, i has o be aken in o accoun hese wo im- po an dimensions: use p e e ences and he use s’ clinical/social p o ile. Pe son-de ined heal hca e goals and p e e ences align p i- ma y and speciali y ca e ocusing on wha ma e s mos o he people. PLWDs and Ca egi e s should be a mo e ac i a ed and engaged in hei ca e when i ocuses on achie ing wha ma e s mos o hem. The main goal o he con en ecommenda ion sys em will be o combine he use p e e ences (based on hei in e es s), he heal h- ca e p o essional ecommenda ions (based on hei p o essional c i e ia) and he Ca egi e s’ opinion (conside ing hei ca e expe i- ence): •A PLWD can selec i s opic p e e ences when egis e ing • A Ca egi e can de ine i s opic p e e ences when egis e ing and also sugges opics o i s ca ed one conside ing his/he speci ic needs • A heal hca e p o ide (medical/social) can de ine opic goals o his managed PLWD/Ca egi e conside ing his/he spe- ci ic needs • O he heal hca e me ada a ela ed o he use p o ile is used o ecommend in e en ions All his p e e ence selec ion is dynamic and can be modi ied in he sys em a any ime, allowing a lexible pe sonaliza ion. 3 RECOMMENDER SYSTEMS BACKGROUND C-MMD pla o m will ake ad an age o ecommende sys em ech- nologies in o de o p o ide PLWD and Ca egi e s wi h a pe son- alized lis o psychosocial in e en ions. This sec ion p o ides an o e iew o he ela ed wo k and me hods ha ha e been in es i- ga ed o de elop he Heal h Recommende Sys em o he C-MMD p ojec . In he mid-nine ies o las cen u y, Recommende Sys ems eme ged as an independen esea ch ield o In o ma ion Re ie al and A - i icial In elligence o add ess he in o ma ion o e load p oblem by using a speci ic ype o in o ma ion il e ing echniques ha a emp s o ecommend in o ma ion i ems (e.g., mo ies, TV, ideos on demand, music, books, news, images, Web pages, esea ch pa- pe s) ha a e likely o be o in e es o he use . In public heal h sys ems, no only he e is an o e load p oblem bu also in o ma ion comes om di e en sou ces and o ma s. Pe sonal heal h eco d sys ems a e mean o cen alise and s anda dise an indi idual’s heal h da a and enhance he da a sha ing among au ho ised p o es- sionals o en i ies. The e o e, ecen ends in esea ch ha e ocused PETRA ’18, June 26–29, 2018, Co u, G eece L. Oli a e al. Table 1: C-MMD In e en ion Types Domain In e en ion In o ma ion and adjus men o diagnosis Demen ia Ad iso s Pos Diagnos ic G oups Signpos ing S ess, anxie y o dep ession managemen Pee Suppo G oups S ess/Anxie y managemen Reminiscence Imp o ing and main aining cogni i e unc ioning Assis i e Technology: ad ice and suppo Cogni i e T aining Heal h and quali y o li e Physical Exe cise he apy Home modi ica ion Fall p e en ion Music The apy on heal h in o ma ion e ie ing and di e en app oaches o Heal h Recommende Sys ems (HRS) can be ound in he li e a u e. An HRS is a specialisa ion o a ecommende sys em, whe e a ecommendable i em o in e es is "a piece o non-con iden ial, sci- en i ically p o en o a leas gene ally accep ed medical in o ma ion, which in i sel is no linked o an indi idual’s medical his o y" [ 32 ]. Possible examples o a ecommended i em a e ood/nu i ion in o - ma ion, a physical ac i i y, a diagnosis, a he apy o a medica ion [ 5 ]. Au ho s in [ 32 ] ha e iden i ied HRS designed o wo a ge publics and pu poses : • diagnos ic o educa ional ool o assis physicians in he decision-making p ocess when ea ing a pa ien • pe sonal heal h ad ising ool o use s, especially ocused on heal hy beha iou al change, engaging use s in o physical ac i i ies o nu i ion based ecommende sys ems The inal objec i e is o empowe pa ien s by educa ing hem abou hei heal h and o e ing means o sel -diagnosis, bu also o p o ide clinicians wi h mo e accu a e in o ma ion abou his/he pa ien s. Finally, heal h ca e p o ide s o insu ances would bene i om his e icien , ailo ed and cos -sa ing se ices [ 27 ]. One o he main challenges o HRS, in compa ison o he ones aced by adi ional ecommende sys ems, is he implica ion o e hical issues when collec ing o p ocessing pe sonal da a and he heal h in o ma ion deli e ed o he end-use . People migh p esen di e en , mul iple condi ions, and each o hem will ha e di e en needs and in e es s. HRS ough o be able o de e mine hese heal h condi ions and which is he meaning ul da a om he pa ien s’ heal h eco d, in o de o p o ide ailo ed, con ex - ela ed, high quali y and us ed ecommenda ions. Nowadays, HRS a e becoming popula due o he inc easing de- mand o nu i ion-based and heal hie li es yle beha iou al ecom- mende s. The In ape sonal Re ospec i e Recommenda ion (IRR, [ 16 ]) is a li es yle change ecommende ha only uses pe sonal his o y and he goal o achie emen . This app oach migh su e om he new use p oblem since ecommenda ions a e based on he e ospec i e his o y and beha iou al pa e ns o an indi idual. MyBeha iou [ 24 ] is a mobile applica ion which acks a use ’s physical ac i i y and die ou ines and p o ides au oma ic heal h eedback. Educa ional HRS is he o he main ield o applica ion. An ex- ample o i is Heal hRecSys [ 26 ], a seman ic-based ecommende which gene a es Medline Plus links ex ac ed om me ada a o selec ed You ube ideos. Au ho s ema k ha he quali y o ecom- menda ions is a ec ed by he seman ic-gap be ween he laype son language and he heal h p o essional. In [ 11 ], au ho s p opose a ecommende sys em ha combines bo h app oaches o de elop a sma phone-based pe sonalized ec- ommende sys em o people wi h dep ession. The sys em will bo h p opose ac i i ies o elie e nega i e emo ions, bu also o e s expe knowledge on how o be awa e o such nega i e emo ions and o help use s lea n how o con ain and change hem. In gene al, HRS a e s ill in an imma u e phase al hough i is an eme ging end in li e a u e. In [ 5 ], au ho s p oposed a amewo k o HRS whe e all he componen s needed o c ea ing a success ul and use ul ecommende sys em a e desc ibed: (i) domain de ini ion, including i ems, con ex , s akeholde s, end-use s o da a a ailabili y, (ii) HRS e alua ion, pa icula ly use accep ance and sa is ac ion, us and p i acy, and communica ion quali y, (iii) beha iou al e alua ions, o how e ec i e was he HRS, (i ) heal h impac , which assesses how beha iou al changes lead o changes in heal h and ( ) e hical conside a ions o he HRS. 3.1 Recommenda ion s a egies In he li e a u e [ 25 ], ecommenda ion s a egies a e classi ied ei- he on he basis o hei knowledge sou ce o o he algo i hmic echnique employed. When classi ied on he basis o he knowledge sou ce, he ollowing wo main s a egies can be iden i ied: con en - based il e ing (CB), which exp esses use in e es s as keywo d- based use p o iles and consis s o ecommending i ems ma ching use p e e ences and i em ea u es; and collabo a i e il e ing (CF), which exp esses use p e e ences as i em a ings and whose ec- ommenda ions a e based on ma ching use s o i ems wi h simila a ing beha iou . Based on he algo i hmic echnique, wo main app oaches can be iden i ied as well: heu is ic-based (also called memo y-based), which employs some heu is ic o mula, such as ec o -based simila i y and co ela ion measu es, o calcula e he ecommenda ion; and model-based, which gene a es he ecom- menda ion using a model lea n by applying some model-building Heal h Recommende Sys em design in he con ex o CAREGIVERSPRO-MMD P ojec PETRA ’18, June 26–29, 2018, Co u, G eece echnique o e he use -i em a ing ma ix. Typically, heu is ic- based sys ems adap be e o changes in use ’s p e e ences bu scale wo se han model-based ones. Heu is ic-based CB [ 1 ] is inspi ed by In o ma ion Re ie al me hods and calcula es he use -i em ma ching based on ec o measu es such as he cosine simila i y. Model-based CB [ 3 ] calcula es he ma ching based on a model o use ’s as es buil by applying Machine Lea ning echniques, such as naï e Bayesian ne wo ks, o ex ca ego iza ion. Heu is ic-based CF employs heu is ic echniques such as a ing co ela ion analysis and is also known as he neighbou hood-based app oach (based on he k-nea es -neighbou algo i hm). Depending on whe he a subse o use s o i ems is chosen o compu e ecommenda ions, wo classes o heu is ic-based CF can be iden i ied: use -based [ 12 ], when he algo i hm ocuses on inding use s simila o he ac i e one o ecommending; and i em-based [ 15 ], when he algo i hm ocuses on inding i ems simila o hose he ac i e use likes. Model- based CF uses he a ings o lea n a compac p edic i e model o ep esen use -i em in e ac ions based on la en ac o s [ 13 ]. In gene al, ecommende sys ems ha e a signi ican dependency on da a. I is possible o dis inguish he ollowing p oblems: (1) Cold-s a p oblem: To p o ide good ecommenda ions i is necessa y o ha e enough eedback om use s (e.g., a ings o p e iously deli e ed in e en ions) o lea n which i ems i hei needs. Fo his eason, new use s ep esen a blank sla e and he sys em does no know how o p o ide pe son- alised esul s since he e a e no elemen s o he decision o do il e ing. This si ua ion is also known as he cold s a p oblem and i also a ec s i ems. I can be usually ound when using collabo a i e il e ing. In his si ua ion, a new i em does no ha e any a ing. (2) Spa si y: Usually, he se o i ems g ows la ge han he se o use s, which implies ha as he sys em ge s olde , he a e age numbe o a ings by i em becomes lowe , since use s end o a e a limi ed numbe o i ems. The e ec o his si ua ion is ha use s become less simila since he e a e less common a ings be ween hem. (3) Gene aliza ion: The gene aliza ion p oblem a ises when he use model is oo gene ic, and i does no allow o make pe - sonalized ecommenda ions bu jus in a gene ic way, hus li le di e ences can be ound om wha is ecommended be ween di e en use s (4) O e specializa ion: Opposed o gene aliza ion, his issue occu s when he ecommende sys em collec s a signi ican amoun o eedback oo ocused on a subse o i ems. In ha case, he sys em may jus ecommend hose i ems ha a e high likely o be al eady known by he use and igno ing he es . (5) Insensi i eness o p e e ence changes: I he use model is no lea n pe iodically, i can u n ou -da ed and do no o e ecommenda ions aligned wi h cu en use ’s p e e ences. In addi ion o his, i is possible o ind Demog aphic-based ap- p oaches (DF), which build s e eo ypes acco ding o use s’ agg e- ga ed pe sonal de ails, hus gene a ing use models om simila use s’ g oups (wi h ega ds demog aphic c i e ia such as age o gende ). Table 2: T ade-o s in ecommenda ion s a egies Limi a ions o Recommenda ion S a egies CB CF DF New use cold-s a p oblem X X New i em cold-s a p oblem X X Spa si y p oblem X X Gene aliza ion p oblem X X O e specializa ion p oblem X Insensi i e o p e e ence changes X X Each algo i hmic app oach has p o ed o ob ain be e esul s o sol e some, bu no all, o hem (see Table 2). In isola ion, CF has p o en o be he mos accu a e app oach in mos domains. Howe e , hese me hods su e om well-known limi a ions ha conside ably a ec hei pe o mance. The mos common solu ion o o e come hese limi a ions in eal applica ions consis s o ap- plying a hyb id s a egy combining CB and CF me hods. Di e en hyb idiza ion s a egies ha e been used in he li e a u e, such as: he weigh ed s a egy, in which he sco e o di e en ecommenda ion componen s a e combined nume ically; he ea u e augmen a ion s a egy, in which he ecommenda ion echnique is used o com- pu e a ea u e o se o ea u es, which is hen pa o he inpu o ano he echnique; and he cascade s a egy, in which ecommen- da ions made by one echnique a e e ined by ano he echnique. Fo ins ance, using DF app oach o deal wi h cold-s a p oblems o new use s, whe e no hing is known abou he use and, a leas , some app oxima ion can be pe o med o o e a use model mo e use ul han jus andomly ecommending. A e wa ds, wi h mo e knowledge abou use s, a mo e e ined use model can be lea n . 4 C-MMD RECOMMENDER COMPONENT This sec ion p esen s an o e iew o he unde aken solu ion ap- p oach o design he Heal h Recommende Sys em. This sys em p o ides wo di e en se ices: a anked lis o in e en ions which is ailo ed o use ’s p e e ences and equi emen s and a lis o po en- ial acquain ances acco ding o hei simila i ies o he gi en use . Al hough bo h unc ionali ies p o ided by he sys em sha e simila componen s and in e aces owa ds da a sou ces, hei objec i es and he na u e o he i ems ha each one handles, di e s. Fo his eason, each pa has been designed di e en ly: hei aining and ecommenda ion p ocesses, al hough basically simila , ha e some sligh di e ences. I is wo h no ing ha he o me has al eady been implemen ed while he la e has no been ye deployed a he momen o w i ing his pape . Ne e heless, we will men ion his second ecommende along he ex as i has a ec ed he o e all design o he sys em. Algo i hmic app oach To de elop he C-MMD HRS we adop ed a hyb id il e ing app oach. I is well known ha hyb id sys ems p o ide be e esul s [ 4 ], con- sequen ly, ha is he chosen app oach. Speci ically, ou app oach consis s o a con en -based algo i hm and a ule-based il e ing which uses a axonomical classi ica ion o in e en ions used by heal h p o essionals ha c ea e hose in e en ions. In his way, we aim o p o ide medical c i e ia o he p ocess o selec ing wha PETRA ’18, June 26–29, 2018, Co u, G eece L. Oli a e al. in e en ions should be e u ned o he use . Ideally, he e is a la ge numbe o possible in e en ions, as many as he di e en kinds o subjec s, a eas o in e es and so ha a use can indica e in i s p o ile. I is also necessa y o no e ha gi en he na u e o he use s, i may be possible ha no oo much eedback can be collec ed in e ms o a ings. This implies ha any eedback e ie ed om use s may be oo spa se, which makes di icul o ind simila use s o p oceed wi h a collabo a i e- il e ing app oach. These easons s eng hen ou app oach o applying a con en -based algo i hm o p edic use ’s sa is ac ion along he use o a ule-based app oach which akes in o conside a ion he opics o in e es chosen by he PLWD, ca egi e and doc o s/social wo ke s. By means o ex- ac ing ea u es ha de ine an in e en ion, i is possible o lea n he p e e ence owa ds each ea u e om use ’s eedback gi en o p e ious in e en ions. Fo ins ance, a use may p o ide posi i e eedback owa ds in e en ions ha a e mainly ideos, explain- ing ips on a speci ic opic o in e es , and nega i e eedback o in e en ions ha a e long ex s. Thus, a con en -based algo i hm can p edic be e sco e o in e en ions ha a e deli e ed in ideo o ma . Besides his, he ecommende sys em also compu es a lis o simila use s o a gi en use . This lis is unde s ood as po en ial acquain ances o iends o ha use , which a e ele an because o he sha ed in e es s hey ha e. Feedback is collec ed h ough he C-MMD pla o m and i s g aphical in e ace. We can dis inguish wo di e en kinds o eedback: (1) Explici : Feedback ha he use has explici ly p o ided, h ough he C-MMD g aphical in e ace, and includes Use ’s p e e - ences and pe sonal de ails. I also includes explici a ings gi en o in e en ions. Explici eedback can become inac- cu a e o ou -da ed since i depends on he use o ac i ely ill in all in o ma ion as well as upda e i . This is especially ue o PLWD use s; hus, we also include he p e e ences sugges ed by ca egi e s and doc o s/social wo ke s wi h ega ds o hei ela ed PLWD (2) Implici : Feedback ha use has no ac i ely p o ided, bu i has been ob ained om he ac ions s/he has ca ied ou and i is cap u ed au oma ically by he pla o m (e.g., in e en ions iewed o sha ed) Ou ini ial design is using bo h kinds o eedback. Speci ically, i is using implici eedback in e ms o in e en ions ha ha e been e ec i ely consumed by he use . This eedback is ob ained h ough he eedback in e ace. Ou sys em is also using explici eedback in e ms o likes/dislikes ha a use may ha e p o ided owa ds in e en ions. I was also conside ed o ask use s o a e using a 1-5 Like scale bu i was disca ded since using his mo e ine-g ained a ing could di icul use expe ience gi en he na u e o hei ill- nesses. Implemen a ion de ails The HRS sys em is composed o he ollowing componen s (see igu e 1): (1) The HRS manage , who akes ca e o o ches a ing he lea n- ing p ocess, combining he esul s o he p edic ions made by bo h ecommende sys ems, apply ha d cons ain s as p e o pos il e ing ules (2) The ecommende sys ems, ha is, he app oaches we e aken o lea n use models and p edic he a ings on non- consumed in e en ions as well as simila use s (3) The adap o s, mean o decouple he in e nal da a models o he ecommende om he da a sou ces models p o ided by o he componen s in he CMMD pla o m, hus con ain- ing po en ial p oblems om changes in hose da a model schemas (4) An endpoin o p o ide he ecommenda ions Figu e 1: HRS a chi ec u e o e iew. Each elemen in he con en -based ecommende s box can consume om di e - en adap o s. The p e ious componen s wo k o line; he use models equi e being upda ed and e-lea n om ime o ime. The upda e e- quency is cu en ly se in a daily basis, bu i may be changed de- pending on he equency ha use s a e he i ems (in e en ions) o change hei p e e ences. The sys em compu es p edic ions and s o es hem o be accessed la e on h ough an API ha publishes hem o each use . This HRS sys em has been implemen ed in Ja a and Cloju e. The o me akes ca e o he engine in e nals while he la e w aps he engine o ake ca e o he public in e ace and da a p epa a ion. We ha e used wo di e en languages so we can use Ja a o ge he ad an age o exis ing lib a ies (e.g., Apache Mahou ) and Cloju e. Using Ja a also acili a es eusing he code mo e easily in di e en scena ios and deploymen s as well as main aining i in he u u e. Cloju e smoo hly in e ope a es wi h Ja a (bo h a e execu ed in he JVM) and acili a es he ask o p o iding a sel - documen ed REST ul API as well as da a cleansing and p epa a ion o he lea ning p ocess pe o med by he Ja a code. Heal h Recommende Sys em design in he con ex o CAREGIVERSPRO-MMD P ojec PETRA ’18, June 26–29, 2018, Co u, G eece As men ioned be o e, he HRS implemen s a con en -based ap- p oach which has been e ined wi h p e- il e ing and pos - ea men p ocesses. The o me emo es hose in e en ions ha a e i el- e an o he use acco ding o he in e es s decla ed by he use , his/he ca egi e and doc o and his/he p o ile (e.g. language, co- mo bidi ies, loca ion). A e wa ds, he sys em p edic s he es ima ed in e es o he use owa ds he po en ial in e en ion candida es ac- co ding o use ’s pe sonal da a (e.g., use ’s de ails), use -gene a ed con en (i.e., a ings, likes, sha es o con en ha was ecommended). The las s ep pos -p ocesses his lis o ank hem acco ding o use ’s in e es s and use ’s ca egi e and doc o indica ions. In his way, he use is ecei ing a lis o in e en ions which a e g ouped acco ding o no only on use ’s in e es s bu also on wha his/he doc o and ca egi e conside mo e use ul o sui able and, wi hin each g oup, anked o he es ima ed p e e ence o he use owa ds hose in e en ions. The lea ning p ocess is pe o med daily and he esul is s o ed o consump ion by he C-MMD pla o m which schedules he appea ance o ecommended in e en ions in use ’s news eed h oughou he day, hus cons an ly p o iding in e es ing in e en- ions o keep use ’s a en ion. 5 CONCLUSIONS In his pape , we ha e ou lined he design o an HRS o he C- MMD pla o m ha is al eady being pilo ed by 600 dyads in ou coun ies. The da a collec ed he nex mon hs will allow us o ine- une he ecommending p ocess and o p o ide esul s on he con en dis ibu ion and consump ion by he di e en s akeholde s. Elde ly people wi h cogni i e impai men eel he e is a g ea need o mo e public awa eness o he disease and mo e suppo o ca egi e s. While ea men s o e e se o hal disease p og ession a e no a ailable o mos o he demen ias, and o he o eseeable u u e, ea men and medica ion o demen ia will emain cen ed en i ely on disease managemen . Slowing down he a e o cogni- i e decline and imp o ing ea ly diagnosis is essen ial in his as i gi es PLWDs he bes chance o main ain hei cogni i e abili y. I also allows ca egi e s and PLWDs o plan o he u u e. C-MMD suppo s PLWDs and Ca egi e s, p o iding ele an in e en ions o hem a he poin o ca e in he communi y, empowe ing hem by building up pa ien s’ capaci y o a be e sel -managemen o he disease and o become ac i e pa ne s in hei own ca e, and o con ibu e o a wide pe spec i e in he heal h ca e sys em. 5.1 Fu u e Wo k Nex de elopmen s eps in he ecommende componen a e headed owa ds he implemen a ion o a second ecommende unc ionali y o p o ide a lis o po en ially in e es ing acquain ances o a gi en use . This will suppo he social ne wo k ha Ca egi e s pla o m is aimed o build. This lis will be based on use s’ simila i ies based on in e es s and in e ac ions wi h p e iously ecommended in e - en ions and also based on use s’ p o iles (e.g., sc eening esul s, como bidi ies, language). ACKNOWLEDGMENTS The wo k p esen ed in his pape has been pe o med unde he Eu opean P ojec CAREGIVERSPRO-MMD[ 6 ], which has ecei ed unding om he Eu opean Union Ho izon 2020 esea ch and inno- a ion p og amme unde g an ag eemen No 690211. P o . Ulises Co és is a membe o Sis ema Nacional de In es i- gado es (SNI-III), México. REFERENCES [1] Jae-wook Ahn, Pe e B usilo sky, Jona han G ady, Daqing He, and Sue Yeon Syn. 2007. Open use p o iles o adap i e news sys ems: help o ha m?. In P oceedings o he 16 h in e na ional con e ence on Wo ld Wide Web. ACM, 11–20. [2] Adminis a ion on Aging" "Alliance o Aging Resea ch. "2012". "T ansla ing Inno- a ion o Impac : E idence-based in e en ions o suppo people wi h Alzheime ’s disease and hei ca egi e s a home and in he communi y". "Me Li e Founda ion". [3] Daniel Billsus and Michael J Pazzani. 1999. A hyb id use model o news s o y classi ica ion. In UM99 Use Modeling. Sp inge , 99–108. [4] Robin Bu ke. 2007. Hyb id web ecommende sys ems. In The adap i e web. Sp inge , 377–408. [5] And é Cale o Valdez, Ma ina Zie le, Ka ien Ve be , Alexande Fel e nig, and And eas Holzinge . 2016. Recommende Sys ems o Heal h In o ma ics: S a e-o - he-A and Fu u e Pe spec i es. Sp inge In e na ional Publishing, Cham, 391–414. h ps://doi.o g/10.1007/978-3-319-50478-0_20 [6] CAREGIVERSPRO-MMD Conso ium. 2018. CAREGIVERSPRO-MMD. (2018). Re ie ed Ma ch 2, 2018 om h ps://ca egi e sp ommd-p ojec .eu [7] G upo de abajo de la Guía de P ác ica Clínica sob e la a ención in eg al a las pe sonas con en e medad de Alzheime y o as demencias. 2010. Guía de P ác ica Clínica sob e la a ención in eg al a las pe sonas con en e medad de Alzheime y o as demencias. Agencia d’In o mació, A aluació i Quali a en Salu de Ca alunya;. [8] Ka ine Pe i p ez e al. 2011. Maladie d’Alzheime e maladies appa en ées : diag- nos ic e p ise en cha ge. Recommenda ion de bonne p a ique. Hau e Au o ié de San é. [9] Na ional Ins i u e o Heal h and Clinical Excellence Social Ca e Ins i u e o Excellence. 2007. A NICE-SCIE Guideline on suppo ing people wi h demen ia and hei ca e s in heal h and social ca e. The B i ish Psychological Socie y and Gaskell. [10] Reinha d Guss and colleagues. 2014. A Guide o Psychosocial In e en ions in Ea ly S ages o Demen ia,. The B i ish Psychological Socie y. [11] Galen Chin-Lun Hung, Pei-Ching Yang, Chen-Yi Wang, and Jung-Hsien Chi- ang. 2015. A Sma phone-Based Pe sonalized Ac i i y Recommende Sys em o Pa ien s wi h Dep ession. In P oceedings o he 5 h EAI In e na ional Con- e ence on Wi eless Mobile Communica ion and Heal hca e (MOBIHEALTH’15). ICST (Ins i u e o Compu e Sciences, Social-In o ma ics and Telecommu- nica ions Enginee ing), ICST, B ussels, Belgium, Belgium, 253–257. h ps: //doi.o g/10.4108/eai.14-10-2015.2261655 [12] Joseph A Kons an, B adley N Mille , Da id Mal z, Jona han L He locke , Lee R Go don, and John Riedl. 1997. G oupLens: applying collabo a i e il e ing o Usene news. Commun. ACM 40, 3 (1997), 77–87. [13] Yehuda Ko en and Robe Bell. 2011. Ad ances in collabo a i e il e ing. In Recommende sys ems handbook. Sp inge , 77–118. [14] S e e Lau iks, Annika Reine smann, Hen ië e Ge alde Van de Roes , FJM Mei- land, Richa d J Da ies, Fe ial Moelae , Mau ice D Mul enna, Ch is D Nugen , and Rose-Ma ie D öes. 2007. Re iew o ICT-based se ices o iden i ied unme needs in people wi h demen ia. Ageing Resea ch Re iews 6, 3 (2007), 223–246. [15] G eg Linden, B en Smi h, and Je emy Yo k. 2003. Amazon. com ecommenda- ions: I em- o-i em collabo a i e il e ing. IEEE In e ne compu ing 7, 1 (2003), 76–80. [16] Be nd Ludwig, F ancesco Ricci, Ze in Yumak, Na a Tin a e , Rong Hu, Pea l Pu, E himios Bo hos, Dimi is Apos olou, G ego is Men zas, Robe G. Fa ell, Ca alina Danis, S ee am Ramak ishnan, Wendy A. Kellogg, Ugan Yasa u , Reza Amini, and Ch is ine L. Lise i. 2012. In ape sonal Re ospec i e Recommen- da ion: Li es yle Change Recommenda ions Using S able Pa e ns o Pe sonal Beha io . [17] Nicos Magla e as, Vasilis Kou kias, Ioanna Chou a da, Dimi ios G. Goulis, A. A amides, D. Adamidis, Geo ge Lou idas, and EA. Balas. 2002. Home ca e deli e y h ough he mobile elecommunica ions pla o m: he Ci izen Heal h Sys em (CHS) pe spec i e. In J Med In o m. 68, 1–3 (2002), 99–111. [18] Esme Moniz-Cook, My a Ve nooij-Dassen, Bob Woods, Ma in O ell, and In e dem Ne wo k. 2011. Psychosocial in e en ions in demen- ia ca e esea ch: The INTERDEM mani es o. Aging & Men al Heal h 15, 3 (2011), 283–290. h ps://doi.o g/10.1080/13607863.2010.543665 a Xi :h ps://doi.o g/10.1080/13607863.2010.543665 PMID: 21491215. [19] Salla Muu aiskangas, Ma ja Ha jumaa, Ki sikka Kaipainen, and Miika E mes. 2016. P ocess and E ec s E alua ion o a Digi al Men al Heal h In e en ion a ge ed a Imp o ing Occupa ional Well-being: Lessons F om an In e en ion S udy Wi h Failed adop ion. JMIR Men Heal h 3, 2 (2016). h ps://doi.o g/doi: 10.2196/men al.4465 PETRA ’18, June 26–29, 2018, Co u, G eece L. Oli a e al. [20] OECD. [n. d.]. Skills Ou look 2013: Fi s Resul s om he Su ey o Adul Skills. ([n. d.]). [21] Ins i u e o Medicine. 2015. Psychosocial In e en ions o Men al and Subs ance Use Diso de s: A F amewo k o Es ablishing E idence-Based S anda ds. The Na ional Academies P ess, Washing on, DC. h ps://doi.o g/10.17226/19013 [22] Wo ld Heal h O ganisa ion. 2016. In e na ional Classi ica ion o Heal h In e en- ions. (2016). Re ie ed Ma ch 2, 2018 om h p://www.who.in /classi ica ions/ ichi/en/ [23] Ronald C. Pe e sen, J.C. S e ens, Ma y Ganguli, E ic G. Tangalos, J.L. Cum- mings, and S e en T. DeKosky. 2001. P ac ice pa ame e : Ea ly de ec- ion o demen ia: Mild cogni i e impai men (an e idence-based e iew). Neu ology 56, 9 (2001), 1133–1142. h ps://doi.o g/10.1212/WNL.56.9.1133 a Xi :h p://n.neu ology.o g/con en /56/9/1133. ull.pd [24] Mash iqui Rabbi, Min Hane Aung, Mi Zhang, and Tanzeem Choudhu y. 2015. MyBeha io : au oma ic pe sonalized heal h eedback om use beha io s and p e e ences using sma phones. In UbiComp. [25] F ancesco Ricci, Lio Rokach, and B acha Shapi a. 2011. In oduc ion o ec- ommende sys ems handbook. In Recommende sys ems handbook. Sp inge , 1–35. [26] Ca los Luis Sanchez Bocaneg a, Jose Luis Se illano Ramos, Ca los Rizo, An on Ci i , and Luis Fe nandez-Luque. 2017. Heal hRecSys: A seman ic con en - based ecommende sys em o complemen heal h ideos. BMC Medical In o - ma ics and Decision Making 17, 1 (15 May 2017), 63. h ps://doi.o g/10.1186/ s12911-017-0431-7 [27] Hanna Schä e , San iago Ho s-F aile, Ragha Pa an Ka umu , And é Cale o Valdez, Alan Said, Helma To kamaan, Tom Ulme , and Ch is oph T a ne . 2017. Towa ds Heal h (Awa e) Recommende Sys ems. In P oceedings o he 2017 In e na ional Con e ence on Digi al Heal h (DH ’17). ACM, New Yo k, NY, USA, 157–161. h ps://doi.o g/10.1145/3079452.3079499 [28] Demen ia Ca e Team. 2018. Demen ia Ca e . (2018). Re ie ed Ma ch 1, 2018 om h p://www.demen iaca e .ne / [29] The PLOS Medicine Edi o s. 2013. A Reali y Checkpoin o Mobile Heal h: Th ee Challenges o O e come. PLoS Medicine 10, 2:e1001395 (2013). h ps: //doi.o g/doi:10.1371/jou nal.pmed.1001395 [30] WHO. 2013. Call o inno a i e heal h echnologies. (2013). Re ie ed Ma ch 1, 2018 om h p://www.who.in /eheal h/en/ [31] WHO. 2016. Demen ia Fac shee . (2016). Re ie ed Ma ch 1, 2018 om h p://www.who.in /mediacen e/ ac shee s/ s362/en/ [32] Ma in Wiesne and Daniel P ei e . 2014. Heal h Recommende Sys ems: Con- cep s, Requi emen s, Technical Basics and Challenges. In e na ional Jou nal o En i onmen al Resea ch and Public Heal h 11, 3 (2014), 2580–2607. h ps: //doi.o g/10.3390/ije ph110302580